XAI’s Grok to be Opensourced 🔓

Today’s top AI Highlights:

1. Elon Musk announces that Grok is being opensource

2. AI dolls for South Korea’s aging population

3. Layers in LLMs are more redundant than you expect

4. Create Custom GPTs within Microsoft Copilot

& so much more!

Read time: 3 mins

You must’ve heard about AI girlfriends for tackling loneliness. But AI companions for elderly people?
A Korean-based company released an AI companion doll at the Mobile World Congress, to combat the loneliness experienced by South Korea’s elderly population. Priced at $1800, this robot doll named Hyodol is not just another tech gadget. Beyond its slightly eerie exterior lies a sophisticated AI capable of conversation, health reminders, and even alerting caregivers in case of emergencies, keeping an eye on its users.

Key Highlights:

1. Hyodol is equipped with a mobile communication module, processor, various sensors, and speakers, supported by an LLM capable of holding full conversations.

2. It offers features like touch interaction, voice reminders, health coaching, quizzes, exercise, music, and more, alongside caregiver functionalities like voice messaging and motion detection.

3. Approximately 7,000 Hyodol dolls have been deployed to seniors living alone and dementia patients in a university hospital in Gangwon province.

4. A study conducted by the company shows that over 9,000 users showed a reduction in depression levels from 5.73 to 3.14 and an improvement in medicine intake.

This is a classic case of AI being effectively deployed for caregiving and companionship. With more acceptance of social robots, it can very well cater to an increasingly aging population and the lowering birth rate.

Elon Musk has announced that xAI will opensource Grok in the coming week. This is shortly after sued OpenAI for deviating from its non-profit mission. Considering the timing of everything, we can’t help but think it to be an after-thought and create a buzz, after he faced criticism for the lawsuit.

For a quick recap, here are all the necessary details on Grok:

1. Grok was launched in December 2023, available to X Premium+ subscribers.

2. Having a rebellious streak and a bit of wit, Grok can engage in conversation with a unique twist, tackling bold and challenging questions. It can give progressive answers on social justice, climate change, and transgender identities.

3. The “killer feature” of Grok is its ability to access real-time data from posts on X to answer questions with up-to-the-minute information.

4. Grok is powered by a generative model Grok-1 that shows impressive results on various benchmarks. Grok API is available and will soon have image and audio recognition functionalities.

5. Grok 1.5 is anticipated soon. The update will feature an Analysis tool designed to summarize threads and responses, and support content creators by offering grammar-checking and draft improvement functionalities.

The sheer size of today’s LLMs, boasting billions or even trillions of parameters, presents a huge challenge when considering the heavy hardware demands. Do all layers contribute to the performance or we’re hauling unnecessary weights? Yes. Many layers of LLMs exhibit high similarity, and some layers play a negligible role in network functionality. Here’s a metric, Block Influence (BI), to assess and prune unnecessary redundant layers, while keeping the models remarkably effective.

Key Highlights:

1. BI measures the transformation of hidden states throughout the model, offering a granular look at which layers are essential and which are redundant. It allows for targeted pruning that maintains the core functionality of the model.

2. Dubbed ShortGPTthis method maintains up to 92% of the model’s performance while removing about 25% of its parameters and computation. For example, eliminating the last 10 layers (25%) of the LLaMA 2-13B model only reduced its performance on the MMLU from 55.0 to 52.2.

3. Beyond its immediate benefits, the study also demonstrates that ShortGPT’s approach is orthogonal to quantization techniques. This suggests that combining both methods could significantly reduce the deployment overhead of LLMs, offering a dual pathway to model optimization.

9 Seconds News for High-performing Individuals

Volv is an app for high-performing individuals. Volv curates the most important and viral content across the internet and delivers it in 9-second articles using AI so you can be updated without doom scrolling on social media. It’s been featured on the Apple App Store and read by 60k+ users globally.

Download Volv here!

1. AirDraw: Transform your surroundings into a canvas! Paint in 3D with just your hands, using AirDraw’s augmented reality. Just go through the instructions and you can start drawing. It’s an even more immersive experience on Apple Vision Pro.

2. Copilot GPT Builder: Create custom GPTs for your specific usecase within the Copilot app. You can give it specific instructions and documents for grounding (RAG) chat responses.

3. LlamaGym: Simplifies fine-tuning of LLM agents with online reinforcement learning, making it easier to experiment with agent prompting and hyperparameters across any Gym environment. It provides an abstract Agent class that handles the complexities of LLM conversation context, episode batches, reward assignment, and PPO setup.

4. Runpod: A globally distributed GPU cloud designed to develop, train, and scale AI applications, featuring over 50 template environments for streamlined workflow and efficient model training. It supports scaling from zero to millions of inference requests with serverless endpoints.

😍 Enjoying so far, TWEET NOW to share with your friends!

1. …even when the competitor’s [AI] chips are free, it’s not cheap enough ~ Jensen Huang

2. Wait a minute, last week it was a trillion dollars and this week it’s free ? ~ Jim Keller

3. I suspect many AI projects will end up in ruin because its developers are just muddling around getting their dopamine hits from the deluge of micro-events about AI. They focus only on the trees but can’t see the forest! ~ Carlos E. Perez

That’s all for today!

See you tomorrow with more such AI-filled content. Don’t forget to subscribe and give your feedback below 👇

⚡️ Follow me on Twitter @Saboo_Shubham for lightning-fast AI updates and never miss what’s trending!!

PS: I curate this AI newsletter every day for FREE, your support is what keeps me going. If you find value in what you read, share it with your friends by clicking the share button below!

Share Unwind AI